@ljoukov/llm 7.0.2 → 7.0.3

package/README.md

@@ -491,6 +491,18 @@ const { value } = await generateJson({
 console.log(value.ok, value.message);
 ```
 
+For Gemini calls that need a specific thinking budget, pass `thinkingBudget`. This overrides the model default and
+`thinkingLevel` for Gemini requests:
+
+```ts
+const { value } = await generateJson({
+  model: "gemini-flash-latest",
+  input: "Return a JSON object with ok=true and message='hello'.",
+  schema,
+  thinkingBudget: 0,
+});
+```
+
 ### Streaming JSON outputs
 
 Use `streamJson()` to stream thought deltas and get best-effort partial JSON snapshots while the model is still

package/dist/index.cjs

@@ -7258,10 +7258,14 @@ function toGemini25ProThinkingBudget(thinkingLevel) {
       return 32768;
   }
 }
-function resolveGeminiThinkingConfig(modelId, thinkingLevel) {
+function resolveGeminiThinkingConfig(modelId, thinkingLevel, thinkingBudget) {
   if (isGeminiImageModelId(modelId) || modelId === "gemini-flash-lite-latest") {
     return void 0;
   }
+  if (thinkingBudget !== void 0) {
+    const normalizedBudget = Math.max(0, Math.floor(thinkingBudget));
+    return normalizedBudget === 0 ? { thinkingBudget: 0 } : { includeThoughts: true, thinkingBudget: normalizedBudget };
+  }
   if (thinkingLevel) {
     if (modelId === "gemini-2.5-pro") {
       return {
@@ -7637,6 +7641,7 @@ function startLlmCallLoggerFromContents(options) {
       ...options.request.imageAspectRatio ? { imageAspectRatio: options.request.imageAspectRatio } : {},
       ...options.request.imageSize ? { imageSize: options.request.imageSize } : {},
       ...options.request.thinkingLevel ? { thinkingLevel: options.request.thinkingLevel } : {},
+      ...options.request.thinkingBudget !== void 0 ? { thinkingBudget: options.request.thinkingBudget } : {},
       ...options.request.mediaResolution ? { mediaResolution: options.request.mediaResolution } : {},
       ...options.request.openAiTextFormat ? { openAiTextFormat: sanitiseLogValue(options.request.openAiTextFormat) } : {},
       ...getCurrentToolCallContext() ? { toolContext: getCurrentToolCallContext() } : {}
@@ -7938,7 +7943,11 @@ async function runTextCall(params) {
             })
           )
         );
-        const thinkingConfig = resolveGeminiThinkingConfig(modelForProvider, request.thinkingLevel);
+        const thinkingConfig = resolveGeminiThinkingConfig(
+          modelForProvider,
+          request.thinkingLevel,
+          request.thinkingBudget
+        );
         const mediaResolution = toGeminiMediaResolution(request.mediaResolution);
         const config = {
           maxOutputTokens: 32e3,
@@ -8245,6 +8254,7 @@ function startJsonStream(request, operation) {
                 responseMimeType: request.responseMimeType ?? "application/json",
                 responseJsonSchema,
                 thinkingLevel: request.thinkingLevel,
+                thinkingBudget: request.thinkingBudget,
                 ...openAiTextFormatForAttempt ? { openAiTextFormat: openAiTextFormatForAttempt } : {},
                 telemetry: false,
                 signal
@@ -9649,7 +9659,11 @@ async function runToolLoop(request) {
           firstModelEventAtMs = Date.now();
         }
       };
-      const thinkingConfig = resolveGeminiThinkingConfig(request.model, request.thinkingLevel);
+      const thinkingConfig = resolveGeminiThinkingConfig(
+        request.model,
+        request.thinkingLevel,
+        request.thinkingBudget
+      );
       const mediaResolution = toGeminiMediaResolution(request.mediaResolution);
       const config = {
         maxOutputTokens: 32e3,